arabic_PP-OCRv6_small_rec

arabic_PP-OCRv6_small_rec

Arabic printed-text recognition model — the first fine-tune of PaddleOCR's PP-OCRv6 recognizer (PP-OCRv6_small_rec, PPLCNetV4 backbone, MultiHead CTC + NRTR) for Arabic. No Arabic PP-OCRv6 recognition model existed before this: PP-OCRv6's unified release covers CJK + 46 Latin scripts only.

This is the headline model of an a→b study: fine-tune the Arabic PP-OCRv5 recognizer (medyas/arabic_PP-OCRv5_mobile_rec, the labelled baseline) and then PP-OCRv6 (this model) on the same task and evaluation protocol. v6 is the shipped model. ⚠️ The v6 arm trained on an upgraded, larger corpus than the published v5 arm (800k vs 500k, plus display fonts and textured backgrounds), so the v5→v6 numbers reflect a backbone and corpus improvement together — not a clean backbone-only ablation (see v5→v6 delta).

• Task: image-to-text (single-line Arabic recognition)
• Algorithm: Backbone=PPLCNetV4, Head=MultiHead (CTCHead + NRTRHead), MultiLoss
• Warm start: PaddlePaddle/PP-OCRv6_small_rec — note this base has no Arabic prior (unlike the v5 arm, which warm-started from an Arabic pretrained model). The Arabic recognition head (~1050 classes) is shape-mismatched against the base head and is randomly re-initialized (non-strict load); only the backbone is warm. Accuracy is ~0 for the first 1–2 epochs by design.
• Corpus: 800k synthetic Arabic printed lines (gen-v2: 26 base fonts + 12 display fonts, textured backgrounds, polarity inversion). This extends the published v1 corpus medyas/arabic-ocr-printed-500k (500k); the 800k v2 superset is not yet on the Hub.
• Selected checkpoint: epoch 11 (lowest CER on a real held-out scanned set)

Evaluation

Numbers are from the peak epoch-11 checkpoint, evaluated on real KITAB-Bench subsets. CER = character error rate, WER = word error rate (lower is better). patsocr is the selection set (real scanned printed text) used to pick the peak epoch; isippt and historyar are held-out and were not used for selection.

¹ The isippt no-mirror CER (10.9%) is computed over the n=341 crops with no decode-mirror ambiguity; the strict columns (13.5% etc.) are over the full n=500 and are the directly leaderboard-comparable number. "Decode-mirror" lines are those containing western digits or bracket pairs, where the visual↔logical reordering is a known decode-time bookkeeping artifact (the model is usually correct; the scorer over-penalizes). We report both so the headline isn't a cherry-pick — see Honest-number disclosure below.

historyar 50.3% is a real domain gap, not a model defect for its target domain. historyar is historical scanned Arabic (degraded paper, old typography, heavy diacritization) that the v1 synthetic corpus does not model. It is the hardest real-print set and is shown as an out-of-domain reference; representative accuracy for modern printed Arabic is the patsocr ≈ 10% (scanned) and isippt ≈ 11–13.5% (documents/presentations) figures. Closing historyar is a v6.1 corpus target (degradation/1-bit/historical-font synthesis).

How good is ~10% CER? Same-set context (KITAB-Bench)

These sets are the OCR subsets of KITAB-Bench (Heakl et al., 2025, arXiv:2502.14949, Table 9), so other systems' scores on the exact same images are published. CER (lower = better):

*Qari hits 0.00 on PATS but collapses to 0.52 on ISI-PPT (overfit-suspect on clean PATS).

This model ranks ≈4th of 12 on patsocr and ≈3rd of 12 on isippt — beating both legacy engines (Tesseract, vanilla PaddleOCR) and most general VLMs (GPT-4o-mini, both Qwen-VLs, AIN, GPT-4o-on-patsocr), and trailing only frontier/specialist systems (Gemini-2.0-Flash, GPT-4o, and Qari-on-patsocr). The vanilla-PaddleOCR row (0.77 / 0.81) shows what fine-tuning on the 500k corpus bought: the same model family goes from near-useless on Arabic to upper-tier. The model's distinguishing strength is cross-set stability (~0.10 / ~0.11) — the strongest competitors each collapse on one set (Azure 0.03→0.98, Qari 0.00→0.52).

Comparability: KITAB-Bench's official scorer (ocr-eval/metrics.py) is diacritic-insensitive plus alef/teh-marbuta/alef-maksura folding + tatweel removal — i.e. more aggressive than this model's normalizer. Our reported CER is therefore a conservative upper bound on the KITAB-comparable figure; re-scoring our predictions through their metrics.py can only hold or lower it.

Honest-number disclosure

Per-set decode-mirror exclusion rates (the n dropped for the no-mirror column): patsocr ~0% (near pure-Arabic), isippt ~32% (n=341 of 500). Because the no-mirror column drops the hardest digit/bracket lines, the strict column is the fair headline; the no-mirror column is a diagnostic, not the marketing number.

v5 → v6 delta (harness-consistent)

v6 (this model) vs the v5 baseline arm — both re-scored under the identical current harness (consolidated eval/contract.py + eval/normalize.py, PaddleOCR v3.7.0; v5 re-scored 2026-06-23, CER lower = better):

v6 improves on every set. One caveat remains (now that the harness is consistent): the corpus differs — v6 trained on the 800k gen-v2 corpus (display fonts + textured backgrounds + polarity), the v5 arm on the 500k v1 corpus. So the gain conflates the PPLCNetV3→V4 backbone change with the corpus upgrade; read it as "the shipped v6 system beats the v5 baseline," not as an isolated backbone ablation (which would require retraining v5 on the 800k corpus, or v6 on the 500k). v6's absolute numbers and the KITAB-Bench comparison above do not depend on this caveat.

Selection curve (patsocr CER, strict)

Epoch 11 was the lowest-CER checkpoint on the real scanned selection set (cold-start v6 needs more epochs than the warm-started v5; the peak-pick self-corrects for over/under-training):

Training

• Hardware: 2× Tesla T4 (Kaggle, free tier), distributed
• Precision: AMP, amp_level=O2, fp16
• Batch: 194 per card
• Epochs: 15 trained, save_epoch_step=1; peak = epoch 11 selected by lowest CER on the real held-out scanned set (patsocr), not on synthetic val
• EMA: off — EMA's fp32 shadow weights are incompatible with the AMP-O2 fp16 model (dtype assert at eval); peak-pick already handles over-training, so EMA is not needed here.
• Multi-session resume: the full run exceeded Kaggle's 12h session cap, so it ran across 2 sessions — per-epoch checkpoints (latest.* + iter_epoch_N.* + optimizer state) were pushed to an HF checkpoint repo each epoch and the second session resumed via Global.checkpoints (not pretrained_model, which would reset the LR schedule). Epoch count and batch size are held fixed across sessions so the cosine LR schedule is continuous.
• Stability: a numpy≥2.0 incompatibility in Paddle's AMP loss-scale NaN path (a (1,)-array float() raise) was patched on disk before launch so transient fp16 inf/nan events skip-and-halve instead of crashing.
• LR: config default 0.0005 (Cosine, warmup_epoch=5) — kept, not per-card-scaled.
• Base config: PaddleOCR configs/rec/PP-OCRv6/PP-OCRv6_small_rec.yml (included as PP-OCRv6_small_rec.yml). max_text_length=25, character_dict_path=ppocrv5_arabic_dict.txt.

Label order

Training labels are stored in VISUAL order (via bidi.get_display). The character dict filename contains the substring arabic (ppocrv5_arabic_dict.txt), which triggers PaddleOCR's CTC-decode pred_reverse path — so tools/infer_rec.py emits predictions in logical (reading) order. Keep the dict filename containing arabic or decoding order will be wrong.

Inference

Do not pip install paddleocr (it pulls torch and breaks the Paddle ABI). Install PaddlePaddle directly and run from a PaddleOCR clone:

git clone https://github.com/PaddlePaddle/PaddleOCR
cd PaddleOCR
# GPU build (CUDA 12.6 wheel index):
pip install paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/stable/cu126/

# download this model's files into <dl>/
hf download medyas/arabic_PP-OCRv6_small_rec --local-dir <dl>

python tools/infer_rec.py \
  -c PP-OCRv6_small_rec.yml \
  -o Global.pretrained_model=<dl>/arabic_PP-OCRv6_small_rec \
     Global.character_dict_path=<dl>/ppocrv5_arabic_dict.txt \
     Global.infer_img=<your_image.png>

(Global.pretrained_model takes the path without the .pdparams suffix. Feed line crops — this is a recognition-only model; pair it with a detector for full-page OCR.)

Deployable inference model (inference/)

The repo also ships a PaddleInference export of the peak (epoch-11) checkpoint under inference/ (inference.json + inference.pdiparams + inference.yml) for serving via PaddleInference / the PaddleOCR predict() API — no PaddleOCR clone or training config needed at runtime.

⚠️ Decode order — important. This model was trained on visual-order labels, so its raw CTC output is in visual (RTL-display) order. tools/infer_rec.py (above) reverses it to logical reading-order automatically because the dict filename contains arabic. The exported inference.yml carries an inline dict (no path), so the bare CTCLabelDecode postprocess does not reverse. To recover logical reading-order from the inference model, apply pred_reverse to its output:

import re
def pred_reverse(pred):  # from PaddleOCR rec_postprocess, for "arabic"-class dicts
    out, cur = [], ""
    for c in pred:
        if not re.search(r"[a-zA-Z0-9 :*./%+-]", c):
            if cur: out.append(cur)
            out.append(c); cur = ""
        else:
            cur += c
    if cur: out.append(cur)
    return "".join(out[::-1])
# logical_text = pred_reverse(model_visual_output)

For guaranteed logical output with zero post-processing, use tools/infer_rec.py with the .pdparams checkpoint (above) — that is exactly how the reported CERs were computed.

Corpus notes

Synthetic Arabic printed lines rendered via libraqm / RAQM (26 base fonts + 12 display fonts: Naskh / Kufi / Ruqaa / clean-sans families; Nastaliq / Diwani / Thuluth deliberately excluded as CTC-unsafe). Text from Wikipedia vocabulary plus ~15% general numerics/dates. Normalization: NFKC, tashkeel stripped (v1), western digits. Real-scan augmentation: blur, gaussian + salt-and-pepper noise, brightness drop, JPEG recompression (q35–70), rotation ±3°, plus textured backgrounds and polarity inversion.

Known gaps (v1 corpus): historical/degraded typography (see historyar); heavy native diacritization. Both are v6.1 corpus targets.

License

Apache-2.0 (inherited from PaddleOCR). The training corpus is synthetic.

Files

• arabic_PP-OCRv6_small_rec.pdparams — peak weights (epoch 11)
• arabic_PP-OCRv6_small_rec.states — training metadata (best metric / epoch)
• ppocrv5_arabic_dict.txt — character dictionary (filename must contain arabic)
• summary_arabic_PP-OCRv6_small_rec.json — full eval summary (selection curve + headline)
• PP-OCRv6_small_rec.yml — base training config (for reproducibility)
• inference/ — deployable PaddleInference export (inference.json / .pdiparams / .yml; see Deployable inference model)